All NaNs Output in `ECCC_downscaling_finetune`

[durbank]

First, thank you guys so much for your work on this! I'm really excited about the possibilities offered by this foundation model.

I am trying to re-implement your eccc_downscaling_finetune example notebook. I've made the relevant updates to configs, regenerated the indices and scalers, etc. Everything runs fine, but after the train_model step, all returned values are NaNs. This is true using either the UNET or CONV model architecture. I've made sure that none of the inputs have missing values. After digging into the intermediary model steps some, I believe I have isolated which step these NaNs are introduced. As near as I can tell, this appears to happen in the Conv2d portion of the final PatchEmbed layer (where batch tensors have shape [1, 512, 32, 32]. I'm at something of a loss, however, as to how to address the problem from here. Would you be able to shed any light on what is going wrong here, and how I might fix it? Thanks in advance for any support/guidance you are able to provide!